Pith. sign in

REVIEW 3 major objections 1 minor 270 cited by

No Language Left Behind: Scaling Human-Centered Machine Translation

T0 review · 3 major / 1 minor · reviewed 2026-05-12 · grok-4.3

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A sparsely gated mixture of experts model trained on mined low-resource data achieves 44% relative BLEU improvement in translating 200 languages.

desk verdict NLLB scales MT to 200 languages with MoE, new low-resource mining, human interviews, and open release, but the 44% BLEU claim needs ablations to separate data from model gains. read the letter →

arxiv 2207.04672 v3 pith:DEJPA3Z5 submitted 2022-07-11 cs.CL cs.AI

NLLB Team , Marta R. Costa-jussà , James Cross , Onur Çelebi , Maha Elbayad , Kenneth Heafield , Kevin Heffernan , Elahe Kalbassi
show 31 more authors
Janice Lam Daniel Licht Jean Maillard Anna Sun Skyler Wang Guillaume Wenzek Al Youngblood Bapi Akula Loic Barrault Gabriel Mejia Gonzalez Prangthip Hansanti John Hoffman Semarley Jarrett Kaushik Ram Sadagopan Dirk Rowe Shannon Spruit Chau Tran Pierre Andrews Necip Fazil Ayan Shruti Bhosale Sergey Edunov Angela Fan Cynthia Gao Vedanuj Goswami Francisco Guzmán Philipp Koehn Alexandre Mourachko Christophe Ropers Safiyyah Saleem Holger Schwenk Jeff Wang (NLLB Team)
This is my paper · ORCID
classification cs.CLcs.AI
keywords machinetranslationlow-resourcelanguagesmultilingualmodelsmixtureofexpertsdataminingsafetyuniversal
checked against Cost.FunctionalEquationFoundation.HierarchyEmergence
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper aims to extend high-quality machine translation to the great majority of languages that current systems handle poorly. Authors first spoke with native speakers to identify real needs, then built new data collection methods and trained a large conditional-compute model on thousands of language pairs. They added training changes to limit overfitting and measured results with human raters on a new benchmark covering all 200 languages plus a separate toxicity check. If the gains hold, far more language pairs become usable in practice and a path opens toward translation support that does not exclude most of the world's languages.

What carries the argument

The sparsely gated mixture of experts architecture, which routes each input to a small subset of experts, combined with tailored data mining that targets low-resource languages.

What would settle it

An independent human evaluation on a fresh set of low-resource sentence pairs that finds no relative BLEU gain or that finds higher rates of toxic outputs.

Watch

Extended reading notes

Core claim

A conditional compute model based on sparsely gated mixture of experts, trained with new data mining techniques for low-resource languages and with added safeguards against overfitting, raises BLEU scores by 44 percent relative to prior state-of-the-art while passing human quality and toxicity evaluations across more than 40,000 translation directions.

Load-bearing premise

The new data mining and model changes produce genuinely better and safer translations instead of merely matching the new benchmark or the particular human raters used.

Editorial extensions

If this is right

  • Thousands of previously unsupported translation directions become accurate enough for everyday use.
  • A combined human-quality and toxicity benchmark becomes a standard way to judge multilingual systems.
  • Releasing the model and mined data lets others add still more languages without starting from scratch.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Conditional routing of computation could be applied to scale speech recognition or summarization to the same wide language set.
  • Gathering direct input from native speakers before model design offers a repeatable way to keep other multilingual tools grounded in actual user needs.
  • Further increases in model size and data coverage could test whether the same techniques eventually support reliable translation even for languages with almost no written data.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 1 minor

Summary. The manuscript presents the No Language Left Behind (NLLB) project, which develops machine translation systems supporting 200 languages with emphasis on low-resource directions. It begins with interviews of native speakers, introduces novel data-mining techniques for low-resource data, trains a Sparsely Gated Mixture-of-Experts model augmented with architectural and training changes to mitigate overfitting across thousands of tasks, and evaluates performance on the human-translated Flores-200 benchmark. The central empirical claim is a 44% relative BLEU improvement over prior state-of-the-art, supported by large-scale human evaluation and a new toxicity benchmark covering all 200 languages; all models, code, and data-mining pipelines are open-sourced.

Significance. If the reported gains prove robust and attributable to the proposed methods rather than data or evaluation artifacts, the work would constitute a substantial advance toward inclusive, high-coverage MT. Strengths include the scale of human evaluation, the introduction of a toxicity benchmark for safety assessment across all languages, the human-centered framing via speaker interviews, and the commitment to open-sourcing. These elements provide concrete resources that could accelerate follow-on research on low-resource translation.

major comments (3)
  1. [Abstract] Abstract: the headline claim of a 44% relative BLEU improvement is presented without any ablation that isolates the contribution of the Sparsely Gated MoE architecture and anti-overfitting training changes from the novel data-mining pipeline. Because the test set (Flores-200) is human-translated and the training data are mined from the same broad web sources, the numerical gain cannot yet be confidently attributed to the model innovations rather than improved data quality or distributional overlap.
  2. [Abstract] Abstract (evaluation paragraph): the aggregate 44% BLEU figure is reported over >40,000 directions, yet no per-language variance, confidence intervals, or statistical significance tests for the relative improvement are referenced. Without these controls it is impossible to determine whether the headline number is driven by a small number of high-resource directions or reflects consistent gains on the low-resource languages that motivate the work.
  3. [Abstract] Abstract: the toxicity benchmark is described as covering all Flores-200 languages and is used to assess translation safety, but the abstract supplies no information on the toxicity classifier, annotation protocol, or decision thresholds. This omission is load-bearing for the safety claims that accompany the performance numbers.
minor comments (1)
  1. [Abstract] The abstract introduces the model as a 'conditional compute model based on Sparsely Gated Mixture of Experts' without immediately clarifying the relationship between the two phrases; a single sentence linking the terms would improve readability.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive comments on the abstract of our NLLB manuscript. We appreciate the recognition of the project's scale, human evaluation, toxicity benchmark, and open-sourcing. All major comments concern the abstract, which we will revise for greater precision and self-containment while preserving its summary nature. We respond point by point below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the headline claim of a 44% relative BLEU improvement is presented without any ablation that isolates the contribution of the Sparsely Gated MoE architecture and anti-overfitting training changes from the novel data-mining pipeline. Because the test set (Flores-200) is human-translated and the training data are mined from the same broad web sources, the numerical gain cannot yet be confidently attributed to the model innovations rather than improved data quality or distributional overlap.

    Authors: We agree that stronger isolation of contributions would be valuable. The manuscript presents the data-mining pipeline and the Sparsely Gated MoE model (with anti-overfitting changes) as complementary elements developed for the 200-language setting, with the 44% gain measured against prior SOTA systems lacking both. The full text contains separate sections detailing each and direct comparisons to prior work. We will revise the abstract to explicitly state that the reported improvement results from the integrated pipeline and to direct readers to the relevant sections for component-wise analysis. Adding exhaustive new ablations at this scale would require substantial additional compute; we therefore treat this as a partial revision focused on abstract clarity. revision: partial

  2. Referee: [Abstract] Abstract (evaluation paragraph): the aggregate 44% BLEU figure is reported over >40,000 directions, yet no per-language variance, confidence intervals, or statistical significance tests for the relative improvement are referenced. Without these controls it is impossible to determine whether the headline number is driven by a small number of high-resource directions or reflects consistent gains on the low-resource languages that motivate the work.

    Authors: The full manuscript and appendices report per-language BLEU scores, variance across directions, and human evaluation results demonstrating that gains are largest and most consistent for low-resource languages. Statistical support comes from the scale of the Flores-200 human evaluations. We will revise the abstract to note that the aggregate figure reflects consistent improvements on low-resource directions, as validated by the detailed per-direction and human assessments presented in the body of the paper. revision: yes

  3. Referee: [Abstract] Abstract: the toxicity benchmark is described as covering all Flores-200 languages and is used to assess translation safety, but the abstract supplies no information on the toxicity classifier, annotation protocol, or decision thresholds. This omission is load-bearing for the safety claims that accompany the performance numbers.

    Authors: We agree the abstract should be more self-contained on this point. The manuscript details a multilingual toxicity classifier fine-tuned on human-annotated data collected from native speakers for each language, together with the annotation protocol and thresholds calibrated via human validation. We will revise the abstract to include a concise description of the toxicity benchmark methodology, noting the classifier, native-speaker annotation, and coverage of all 200 languages. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: empirical benchmark result with no self-referential derivations

full rationale

The paper reports an engineering achievement: a sparsely-gated MoE model trained on newly mined low-resource data, with listed anti-overfitting changes, evaluated on the human-translated Flores-200 benchmark. The 44% relative BLEU figure is a direct measurement on held-out test data rather than a prediction derived from fitted parameters or prior self-citations. No equations, uniqueness theorems, or ansatzes are invoked that reduce the claimed improvement to the inputs by construction. The work is self-contained as an empirical report; any self-citations are incidental and not load-bearing for the central numerical claim.

Assumptions & free parameters 2 free parameters · 2 assumptions · 0 invented entities

The performance claim rests on the assumption that the mined data is representative and that the chosen MoE hyperparameters and regularization successfully prevent overfitting on low-resource pairs; these are not derived from first principles.

free parameters (2)
  • number of experts and expert capacity
    Hyperparameters of the sparsely gated MoE that control compute allocation across thousands of language pairs.
  • regularization coefficients and training schedule
    Values chosen to counteract overfitting when training on many low-resource tasks simultaneously.
assumptions (2)
  • domain assumption Mined parallel data for low-resource languages is sufficiently clean and representative for supervised training.
    Invoked when claiming that novel mining techniques narrow the performance gap.
  • domain assumption Human raters and the toxicity classifier provide reliable safety signals across all 200 languages.
    Required for the claim that translations are both high-quality and safe.

how reviews work

0 comments
Cite this review

Pith. "Pith review of No Language Left Behind: Scaling Human-Centered Machine Translation." pith.science (2026). https://pith.science/paper/DEJPA3Z5

@misc{pith2026220704672,
  author       = {Pith},
  title        = {Pith review of: No Language Left Behind: Scaling Human-Centered Machine Translation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DEJPA3Z5}},
  note         = {Machine review of arXiv:2207.04672}
}
read the original abstract

Driven by the goal of eradicating language barriers on a global scale, machine translation has solidified itself as a key focus of artificial intelligence research today. However, such efforts have coalesced around a small subset of languages, leaving behind the vast majority of mostly low-resource languages. What does it take to break the 200 language barrier while ensuring safe, high quality results, all while keeping ethical considerations in mind? In No Language Left Behind, we took on this challenge by first contextualizing the need for low-resource language translation support through exploratory interviews with native speakers. Then, we created datasets and models aimed at narrowing the performance gap between low and high-resource languages. More specifically, we developed a conditional compute model based on Sparsely Gated Mixture of Experts that is trained on data obtained with novel and effective data mining techniques tailored for low-resource languages. We propose multiple architectural and training improvements to counteract overfitting while training on thousands of tasks. Critically, we evaluated the performance of over 40,000 different translation directions using a human-translated benchmark, Flores-200, and combined human evaluation with a novel toxicity benchmark covering all languages in Flores-200 to assess translation safety. Our model achieves an improvement of 44% BLEU relative to the previous state-of-the-art, laying important groundwork towards realizing a universal translation system. Finally, we open source all contributions described in this work, accessible at https://github.com/facebookresearch/fairseq/tree/nllb.

Discussion (0). Continue with ORCID to comment.

Forward citations

Showing 60 of 270 Pith papers that cite this

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. See all 270 Pith citations

  1. DataComp-VLM: Improved Open Datasets for Vision-Language Models

    cs.CV 2026-06 unverdicted novelty 8.0 of 10

    DataComp-VLM benchmark shows instruction-heavy data mixtures outperform caption-heavy ones for VLM training, with DCVLM-Baseline reaching 63.6% on 33 tasks using 200B tokens, +5.4pp over FineVision.

  2. Massive Activations in Hybrid Linear Attention Large Language Models: Pre-Attention Spikes and Inter-Spike Plateaus

    cs.CL 2026-08 conditional novelty 7.0 of 10

    Massive activations in hybrid linear attention LLMs form pre-attention spikes before full attention layers and inter-spike plateaus that grow with full attention density.

  3. Actions Speak Louder than Words: Measuring Cross-Lingual Policy Retention in Tool-Using Agents

    cs.CL 2026-08 conditional novelty 7.0 of 10

    With five confounds corrected, four frontier tool-using models keep 71-73% of their action-policy consistency when the language changes, and the apparent small-model ordering is largely a chance-floor artifact.

  4. Strengthening Target-Language Features: SAE-Based Steering for Multilingual Inference

    cs.CL 2026-08 conditional novelty 7.0 of 10

    Steering the top SAE features selected by target-language contrast improves multilingual benchmark accuracy in Gemma-3-12B-it without parameter updates.

  5. ALEE: Any-Language Evaluation of Embeddings via English-Centric Minimal Pairs

    cs.CL 2026-06 unverdicted novelty 7.0 of 10

    ALEE generates AMR-based English minimal pairs with fine-grained semantic shifts, translates them, and evaluates embedding models on 275+ languages to expose cross-lingual gaps linked to training data and tokenization.

  6. MMed-Bench-IR: A Heterogeneous Benchmark for Multilingual Medical Information Retrieval

    cs.CL 2026-06 unverdicted novelty 7.0 of 10

    MMed-Bench-IR is a new heterogeneous benchmark spanning 6 languages and three non-overlapping tasks that exposes severe cross-lingual drops in biomedical retrieval performance.

  7. A Komi-Yazva--Russian Parallel Corpus and Evaluation Protocol for Zero- and Few-Shot LLM Translation

    cs.CL 2026-06 unverdicted novelty 7.0 of 10

    Releases a 457-sentence Komi-Yazva--Russian parallel corpus and shows that retrieval-based few-shot prompting improves LLM translation over zero-shot in this low-resource setting, with performance varying by model and metric.

  8. Multilingual Coreference Resolution via Cycle-Consistent Machine Translation

    cs.CL 2026-06 unverdicted novelty 7.0 of 10

    A cycle-consistent MT pipeline generates and similarity-weights training data for coreference resolution, producing gains on four low-resource languages and enabling the task where no corpora existed.

  9. Multilingual Idioms in Sentences and Conversations Across High-, Medium-, and Low-Resource Languages

    cs.CL 2026-06 unverdicted novelty 7.0 of 10

    MIDI is a new multilingual idiom dataset with sentence and conversational contexts; benchmarking reveals worse performance in low-resource languages and on literal vs. figurative uses.

  10. BrahmicTokenizer-131K: An Indic-Capable Drop-In Replacement for o200k_base

    cs.CL 2026-05 unverdicted novelty 7.0 of 10

    BrahmicTokenizer-131K is a 131K-vocab tokenizer constructed via script-prune crop and linear-programming retrofit to o200k_base, achieving 26.7% fewer tokens on Indic text while matching o200k_base on English fertilit...

  11. FTibSuite: A Comprehensive Resource Suite for Tibetan Vision-Language Modeling

    cs.CV 2026-05 unverdicted novelty 7.0 of 10

    FTibSuite provides human-verified multimodal corpora, Tibetan-adapted benchmarks with quality controls, and a baseline VLM showing gains on tasks like MMBench while preserving Chinese capabilities.

  12. A Data-Efficient Path to Multilingual LLMs: Language Expansion via Post-training PARAM$\Delta$ Integration into Upcycled MoE

    cs.CL 2026-05 unverdicted novelty 7.0 of 10

    PARAMΔ upcycles dense models to MoE for per-language experts and grafts post-training deltas to enable data-efficient language expansion while preserving original capabilities.

  13. Knowledge Beyond Language: Bridging the Gap in Multilingual Machine Unlearning Evaluation

    cs.CL 2026-05 unverdicted novelty 7.0 of 10

    New metrics KSS and KPS are introduced to evaluate multilingual machine unlearning quality and cross-language consistency in LLMs, addressing limitations of single-language evaluation protocols.

  14. One Model to Translate Them All? A Journey to Mount Doom for Multilingual Model Merging

    cs.CL 2026-04 unverdicted novelty 7.0 of 10

    Merging fine-tuned models for multilingual translation fails because fine-tuning redistributes language-specific neurons rather than sharpening them, increasing representational divergence in output-generating layers.

  15. Beyond Literal Mapping: Benchmarking and Improving Non-Literal Translation Evaluation

    cs.CL 2026-01 conditional novelty 7.0 of 10

    MENT benchmark plus RATE agentic evaluator raise combined system- and segment-level correlation with human judgments by at least 3.2 points over prior MT metrics and LLM judges.

  16. When the Gold Standard Isn't Necessarily Standard: Challenges of Evaluating the Translation of User-Generated Content

    cs.CL 2025-12 unverdicted novelty 7.0 of 10

    Translation guidelines for user-generated content vary widely across datasets, and prompt-aligned guidelines can shift LLM translation scores by several points, so evaluation must account for them.

  17. Hearing to Translate: The Effectiveness of Speech Modality Integration into LLMs

    cs.CL 2025-12 unverdicted novelty 7.0 of 10

    Cascaded systems remain the most reliable for speech translation overall, but recent SpeechLLMs match or outperform them in many conditions while standalone speech models lag.

  18. Swivuriso: The South African Next Voices Multilingual Speech Dataset

    cs.CL 2025-12 conditional novelty 7.0 of 10

    Swivuriso provides a 3,000-hour, seven-language, scripted-and-unscripted South African speech corpus with domain coverage in agriculture, healthcare, and general topics.

  19. How to Evaluate Speech Translation with Source-Aware Neural MT Metrics

    cs.CL 2025-11 unverdicted novelty 7.0 of 10

    Source-aware MT metrics adapted to speech translation via ASR transcripts or back-translations as audio proxies, plus a new cross-lingual re-segmentation algorithm, improve correlation with human judgments over refere...

  20. FLEXITOKENS: Flexible Tokenization for Evolving Language Models

    cs.CL 2025-07 unverdicted novelty 7.0 of 10

    FLEXITOKENS replaces rigid subword tokenizers and fixed-compression auxiliary losses with a simplified boundary-prediction objective in byte-level models, yielding lower over-fragmentation and up to 10-point gains on ...

  21. Translate With Care: Addressing Gender Bias, Neutrality, and Reasoning in Large Language Model Translations

    cs.CL 2025-05 conditional novelty 7.0 of 10

    A new genderless-to-English benchmark shows that fine-tuning mBART-50 on carefully curated examples cuts gender stereotyping and pronoun-reasoning errors, beating larger proprietary systems on that benchmark.

  22. Memorization Inheritance in Sequence-Level Knowledge Distillation for Neural Machine Translation

    cs.CL 2025-02 conditional novelty 7.0 of 10

    Sequence-level knowledge-distilled NMT students memorize more of the original corpus and hallucinate more than same-size baselines trained directly on that corpus, despite never seeing it.

  23. Evaluating the Capabilities of Large Language Models for Multi-label Emotion Understanding

    cs.CL 2024-12 conditional novelty 7.0 of 10

    A new multi-label emotion benchmark for four Ethiopian languages shows that fine-tuned encoder-only models outperform zero-shot and few-shot large language models, with large gaps between resource-rich and resource-po...

  24. Large Concept Models: Language Modeling in a Sentence Representation Space

    cs.CL 2024-12 conditional novelty 7.0 of 10

    A sentence-level language model trained to autoregressively predict SONAR sentence embeddings can summarize, expand, and generate text in unseen languages.

  25. Improving Language Transfer Capability of Decoder-only Architecture in Multilingual Neural Machine Translation

    cs.CL 2024-12 conditional novelty 7.0 of 10

    A two-stage decoder-only architecture with instruction-level contrastive learning improves zero-shot multilingual translation and closes most of the gap to encoder-decoder models.

  26. Task Prompt Vectors: Effective Initialization through Multi-Task Soft-Prompt Transfer

    cs.CL 2024-08 unverdicted novelty 7.0 of 10

    Task prompt vectors, formed by subtracting random initialization from tuned soft prompts, support low-resource initialization and arithmetic combination across tasks on 12 NLU datasets while remaining independent of i...

  27. Towards Measuring the Representation of Subjective Global Opinions in Language Models

    cs.CL 2023-06 conditional novelty 7.0 of 10

    LLMs default to responses more similar to opinions from the USA and some European and South American countries; prompting for a country shifts alignment but can introduce stereotypes, while translation does not reliab...

  28. Language Models are not Equally Robust to Non-Canonical Tokenization across Languages

    cs.CL 2026-07 conditional novelty 6.5 of 10

    Tokenization invariance does not generalize beyond English: non-canonical segmentations cut multilingual LLM task scores by ~10–24% on average, worse for high-fragmentation languages, and multi-tokenization LoRA mitig...

  29. Build it, Break it, Repeat: Benchmarking and improving LLM-manipulated disinformation detection in social media posts

    cs.CL 2026-08 conditional novelty 6.0 of 10

    An iterative adversarial framework shows that chained back-translation and persona rewrites flip detector labels up to 95% of the time, and a triplet contrastive detector with dynamic anchor switching remains the most robust.

  30. Test-Time Augmentation for LLMs: When Input Diversity Beats Output Diversity at Matched Compute

    cs.LG 2026-08 conditional novelty 6.0 of 10

    For mid-tier LLMs, generating paraphrased versions of an input and majority voting over answers converts inference budget into accuracy more efficiently than self-consistency sampling.

  31. Hidden Language Consistency Phenomena in Reasoning LLMs

    cs.CL 2026-08 conditional novelty 6.0 of 10

    Reasoning models often stop using the requested language as problems get harder, and this language breakdown can make accuracy look better than it is.

  32. Mind the Cap: Output-Budget Regimes Change the Measured Multilingual Reasoning Gap

    cs.CL 2026-08 conditional novelty 6.0 of 10

    The measured native-versus-translate reasoning gap on MGSM depends strongly on the output-token budget, nearly vanishing at saturation and reversing direction under tight caps.

  33. Disentangling Language Modeling and Boundaries

    cs.CL 2026-08 reject novelty 6.0 of 10

    The paper hypothesizes that next-byte and boundary distributions in byte-level LMs can be disentangled, proposes two experiments to test it, but provides no experimental results.

  34. Cross-Lingual Transfer for Machine Translation in Turkic Languages

    cs.CL 2026-07 conditional novelty 6.0 of 10

    Among five Turkic languages, mT5 transfer is strongest for Turkish–Azerbaijani and Kazakh–Kyrgyz, direction and translation target matter, and Latinization helps surface metrics mainly in script-mismatched pairs.

  35. Inspect India Evals: An Open Benchmarking Framework for Evaluating Large Language Models in the Indian Linguistic and Cultural Context

    cs.CL 2026-07 conditional novelty 6.0 of 10

    Inspect India Evals provides six India-specific LLM benchmarks and preliminary scores for five open-weight models, with most performance gaps not statistically significant at n=5.

  36. LatentMT: Machine Translation with Latent Reasoning

    cs.CL 2026-07 conditional novelty 6.0 of 10

    A 2.6B looped language model with per-pair LoRA adapters matches or beats 8B-14B MT systems on 32 language pairs, with recurrent-step gains saturating after the first few steps.

  37. Rate-Utility Frontiers for Language Encodings: Comparing Tokens, Bytes, and Pixels Under Controlled Linguistic Content

    cs.CL 2026-07 accept novelty 6.0 of 10

    No text encoding dominates: pixels preserve surface form, bytes align languages, and tokens classify topics, with the winner depending on capacity, language mix, and compute.

  38. Andha-Dhun: A First Look at Audio Descriptions in Hindi

    cs.CV 2026-07 conditional novelty 6.0 of 10

    The paper introduces Andha-Dhun, the first Hindi audio description dataset, and shows that direct generation from dense captions outperforms translation of English ADs, while machine translation fails to resolve cultu...

  39. Can Dialects Be Steered Like Languages? Sparse Neurons and Distributed Directions in Arabic LLMs

    cs.CL 2026-07 conditional novelty 6.0 of 10

    Arabic dialects are causally steerable in LLMs via sparse LAPE neurons and distributed activation vectors, with vector steering giving more reliable dialect control than neuron rescaling.

  40. The Heterogeneous Safety Impacts of Benign Multilingual Fine-Tuning

    cs.CL 2026-06 unverdicted novelty 6.0 of 10

    Benign multilingual fine-tuning causes language-specific safety drifts with adversarial compliance rates rising up to four-fold, decoupled from capability gains.

  41. Soft Token Alignment for Cross-Lingual Reasoning

    cs.CL 2026-06 unverdicted novelty 6.0 of 10

    SOLAR aligns soft-token probability mixtures across languages in embedding space during SFT and raises multilingual reasoning accuracy by up to 17.7 points over the base model.

  42. The Tatoxa System for Text Detoxification in Low-Resource Languages: The Case of Tatar

    cs.CL 2026-06 unverdicted novelty 6.0 of 10

    Tatoxa outperforms open and commercial LLMs for Tatar text detoxification using a new dataset, with native Tatar training beating cross-lingual transfer from Russian.

  43. SARA: Unlocking Multilingual Knowledge in Mixture-of-Experts via Semantically Anchored Routing Alignment

    cs.CL 2026-06 unverdicted novelty 6.0 of 10

    SARA aligns internal routing distributions in MoE layers to high-resource semantic anchors via symmetric JS divergence, improving low-resource language performance by 0.8-1.2% over standard instruction tuning on Global-MMLU.

  44. Multilingual Hematology Visual Question Answering Dataset

    cs.CV 2026-06 unverdicted novelty 6.0 of 10

    Introduces WBCMor VQA benchmark with 110K bilingual QA pairs for hematology VQA on 20K cell images using existing annotations and a new Urdu dictionary.

  45. MADE: Beyond Scoring via a Multilingual Agentic Diagnosing Engine for Fine-Grained Evaluation Insights

    cs.CL 2026-06 unverdicted novelty 6.0 of 10

    MADE is a new multilingual agentic diagnosing engine that produces higher-quality diagnostic reports (47% better than baseline) on a large-scale evaluation substrate covering 33 model families and 26 languages.

  46. Exploring Adversarial Robustness and Safety Alignment in Multilingual Multi-Modal Large Language Models

    cs.CL 2026-06 unverdicted novelty 6.0 of 10

    Adversarial images transfer across languages in MLLMs while apparent safety in weaker languages stems from comprehension and visual-grounding failures rather than genuine alignment.

  47. Beyond "To whom it may concern": Tailoring Machine Translation to Audience and Intent

    cs.CL 2026-06 unverdicted novelty 6.0 of 10

    Explicit purpose instructions improve LLM translation adaptedness across 50 languages and 8 domains, with larger gains on informal text, while standard metrics often penalize the adapted outputs.

  48. Benchmarking Speech-to-Speech Translation Models

    cs.CL 2026-06 unverdicted novelty 6.0 of 10

    COMPASS is a new reproducible benchmarking framework for S2ST that deploys 46 metrics on 1248 configurations, shows single-metric rankings mislead, reduces to 10 metrics per direction, and finds domain-specific metric...

  49. Translating Classical Poetry into Modern Prose

    cs.CL 2026-06 unverdicted novelty 6.0 of 10

    Six hundred classical Telugu poems with prose translations expose that LLMs mistranslate poetic meaning and that existing automatic metrics miss hallucinated content.

  50. From Outliers to Errors: Auditing Pali-to-English LLM Translations with Multi-Reference Adjudication

    cs.CL 2026-05 unverdicted novelty 6.0 of 10

    A multi-reference audit framework for LLM translations of the Pali Canon uses embedding drift from a human reference centroid to triage candidates for LLM-judge adjudication, showing drift correlates with major error ...

  51. HardMTBench: Stress-Testing Chinese-English Translation on Knowledge-Intensive Domains

    cs.CL 2026-05 unverdicted novelty 6.0 of 10

    HardMTBench is a difficulty-aware benchmark of 20,000 directional test items across 12 domains that widens GEMBA score ranges by a factor of two and reveals domain-specific weaknesses in 22 MT systems.

  52. DEPART: DEcomposing PARiTy across Multilingual LLMs

    cs.CL 2026-05 unverdicted novelty 6.0 of 10

    A Bayesian framework decomposes mLLM variance, showing language features explain 79-92% of language identity variance and that model identity vs. benchmark-model interactions dominate differently for understanding ver...

  53. Extracting Small Translation Specialists from LLMs by Aggressively Pruning Experts

    cs.CL 2026-05 unverdicted novelty 6.0 of 10

    Aggressive expert pruning in MoE LLMs extracts compact translation specialists that retain near-baseline quality after removing up to 75% of experts (or 90% with short SFT).

  54. ICCU: In-Context Continual Unlearning via Pattern-Induced Refusal Rules

    cs.AI 2026-05 unverdicted novelty 6.0 of 10

    ICCU induces pattern-based refusal rules from forget data and applies them in-context at inference to achieve continual unlearning without parameter changes or cross-request interference.

  55. Testing the Deliteralization Hypothesis in Human and Machine Translation

    cs.CL 2026-05 unverdicted novelty 6.0 of 10

    LLMs deliteralize monotonically during self-revision and invert human post-editing triggers, while human translations remain less literal than all tested machine systems.

  56. NASiC: 3D NAND-based CAM-Selected Multibit CIM Architecture for Efficient On-Device Mixture-of-Experts LLM Inference

    cs.AR 2026-05 unverdicted novelty 6.0 of 10

    NASiC fuses CAM-based expert selection and multibit CIM computation in 3D NAND into one cycle for MoE LLM inference, claiming 4-114.8x performance and 3.9-70x energy efficiency gains over prior designs with high accuracy.

  57. Multilingual Steering by Design: Multilingual Sparse Autoencoders and Principled Layer Selection

    cs.CL 2026-05 unverdicted novelty 6.0 of 10

    Multilingual SAEs strengthen cross-lingual representations for reliable steering and an intersection-based rule selects effective layers without exhaustive search.

  58. Direct Translation between Sign Languages

    cs.CL 2026-05 unverdicted novelty 6.0 of 10

    Direct sign-to-sign translation via back-translation synthetic data and joint MBART training outperforms cascaded sign-text-sign baselines on geometric error and BLEU metrics while running 2.3 times faster.

  59. Mix, Don't Tune: Bilingual Pre-Training Outperforms Hyperparameter Search in Data-Constrained Settings

    cs.LG 2026-05 conditional novelty 6.0 of 10

    Mixing auxiliary high-resource language data outperforms hyperparameter tuning in data-constrained bilingual pre-training, with gains equivalent to 2-13 times more unique target data.

  60. ATD-Trans: A Geographically Grounded Japanese-English Travelogue Translation Dataset

    cs.CL 2026-05 conditional novelty 6.0 of 10

    ATD-Trans is a new geographically annotated Japanese-English travelogue dataset that reveals Japanese-enhanced models perform better on geo-entity translation while domestic Japanese locations remain harder to transla...

See all 270 Pith citations

Reference graph

Works this paper leans on

16 extracted references · 16 canonical work pages · cited by 270 Pith papers (see all)

  1. [1]

    URL https://arxiv.org/abs/2110.03036. Benjamin Akera, Jonathan Mukiibi, Lydia Sanyu Naggayi, Claire Babirye, Isaac Owomugisha, Solomon Nsumba, Joyce Nakatumba-Nabende, Engineer Bainomugisha, Ernest Mwebaze, and John Quinn. Machine translation for african languages: Community creation of datasets and models in uganda. In 3rd Workshop on African Natural Lan...

    work page arXiv

  2. [2]

    Farhad Akhbardeh, Arkady Arkhangorodsky, Magdalena Biesialska, Ondřej Bojar, Rajen Chatterjee, Vishrav Chaudhary, Marta R

    URL https://openreview.net/forum?id=BK-z5qzEU-9. Farhad Akhbardeh, Arkady Arkhangorodsky, Magdalena Biesialska, Ondřej Bojar, Rajen Chatterjee, Vishrav Chaudhary, Marta R. Costa-jussa, Cristina España-Bonet, Angela Fan, Christian Federmann, Markus Freitag, Yvette Graham, Roman Grundkiewicz, Barry Haddow, Leonie Harter, Kenneth Heafield, Christopher Homan,...

    work page 2021

  3. [3]

    Fine-grained prediction of reading comprehension from eye movements,

    JMLR.org, 2016. Faisal Alshargi, Shahd Dibas, Sakhar Alkhereyf, Reem Faraj, Basmah Abdulkareem, Sane Yagi, Ouafaa Kacha, Nizar Habash, and Owen Rambow. Morphologically annotated corpora for seven Arabic dialects: Taizi, sanaani, najdi, jordanian, syrian, iraqi and Moroccan. In Proceedings of the Fourth Arabic Natural Language Processing Workshop , 2019. C...

    work page doi:10.18653/v1/ 2016

  4. [4]

    doi: 10.18653/v1/2021.iwslt-1.1

    Association for Computational Linguistics. doi: 10.18653/v1/2021.iwslt-1.1. URL https://aclanthology.org/2021.iwslt-1.1. Patrick Andries. Proposition d’ajout de l’écriture tifinaghe. Organisation internationale de normalisation. Jeu universel des caractères codés sur octets (JUC). ORGANISATION INTERNATIONALE DE NORMALISATION, 2004. Mohd Zeeshan Ansari, M....

    work page doi:10.18653/v1/2021.iwslt-1.1 2021

  5. [5]

    Estimating or Propagating Gradients Through Stochastic Neurons for Conditional Computation

    URL http://arxiv.org/abs/1308.3432. Abhik Bhattacharjee, Tahmid Hasan, Wasi Uddin Ahmad, and Rifat Shahriyar. Banglanlg: Benchmarks and resources for evaluating low-resource natural language generation in bangla. arXiv preprint arXiv:2205.11081 , 2022. Steven Bird. Designing for language revitalisation. In Gilles Adda, Khalid Choukri, Irm- garda Kasinskai...

    work page Pith review arXiv doi:10.18653/v1/2020.acl-main.485 2022

  6. [6]

    Kevin Degila, Godson Kalipe, Jamiil Touré Ali, and Momboladji Balogoun

    ISBN 9788858113622. Kevin Degila, Godson Kalipe, Jamiil Touré Ali, and Momboladji Balogoun. Parallel text dataset for Neural Machine Translation (French -> Fongbe, French -> Ewe), November

    work page

  7. [7]

    Stefano Demichelis and Jorgen W Weibull

    URL https://doi.org/10.5281/zenodo.4266935. Stefano Demichelis and Jorgen W Weibull. Language, meaning, and games: A model of communication, coordination, and evolution. American Economic Review, 98(4):1292–1311, 2008. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language under...

    work page doi:10.5281/zenodo.4266935 2008

  8. [8]

    Jukebox: A Generative Model for Music

    URL https://aclanthology.org/N19-1423. Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, and Ilya Sutskever. Jukebox: A generative model for music. arXiv preprint arXiv:2005.00341 , 2020. Jesse Dodge, Taylor Prewitt, Remi Tachet Des Combes, Erika Odmark, Roy Schwartz, Emma Strubell, Alexandra Sasha Luccioni, Noah A Smith, Nicole...

    work page Pith review arXiv 2005

Show all 16 references
  1. [9]

    low-resource

    Technical report, World Health Organization, May 2017. Nan Du, Yanping Huang, Andrew M. Dai, Simon Tong, Dmitry Lepikhin, Yuanzhong Xu, Maxim Krikun, Yanqi Zhou, Adams Wei Yu, Orhan Firat, Barret Zoph, Liam Fedus, Maarten Bosma, Zongwei Zhou, Tao Wang, Yu Emma Wang, Kellie Web...

    2017 doi

  2. [10]

    multilingual

    Association for Computational Linguistics. doi: 10.18653/v1/W18-6319. URL https://aclanthology.org/W18-6319. Manasa Prasad, Theresa Breiner, and Daan van Esch. Mining training data for language modeling across the world’s languages. In SLTU, pages 61–65, 2018. Ivan Provilkov, ...

    2018 doi

  3. [11]

    Shashi Shekhar, Dilip Kumar Sharma, and MM Sufyan Beg

    URL https://openreview.net/pdf?id=B1ckMDqlg. Shashi Shekhar, Dilip Kumar Sharma, and MM Sufyan Beg. Language identification framework in code-mixed social media text based on quantum lstm —the word belongs to which language? Modern Physics Letters B , 34(06):2050086, 2020. Adi...

    2020 doi

  4. [12]

    doi: 10.18653/v1/2021.eacl-main.163

    Association for Computational Linguistics. doi: 10.18653/v1/2021.eacl-main.163. URL https://aclanthology.org/2021.eacl-main.163. Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V. Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, Jeff Klin...

    2021 doi arXiv

  5. [13]

    civilizing

    Association for Computational Linguistics. doi: 10.18653/v1/D16-1163. URL https://aclanthology.org/D16-1163. Barret Zoph, Irwan Bello, Sameer Kumar, Nan Du, Yanping Huang, Jeff Dean, Noam Shazeer, and William Fedus. Designing effective sparse expert models. arXiv preprint arXi...

    2022 doi

  6. [14]

    (2018) 3 4 AI4D Degila et al

    https://github.com/facebookresearch/fairseq/tree/nllb/data 175 Corpus Name Citation # Directions # Languages AAU Ethiopian Languages Abate et al. (2018) 3 4 AI4D Degila et al. (2020); Siminyu et al. (2021) 3 5 DGT Tiedemann (2012) 94 24 ECB Tiedemann (2012) 74 19 EMEA Tiedeman...

    2018

  7. [15]

    We compare non-English-centric performance in this table

    4-phase curriculum : (a) Step 0−170k: https://github.com/facebookresearch/fairseq/tree/nllb/examples/ nllb/modeling/scripts/flores200/final_lang_pairs_cl3.txt (b) Step 170k−230k: https://github.com/facebookresearch/fairseq/tree/nllb/examples/ nllb/modeling/scripts/flores200/fi...

  8. [16]

    Naive 2-phase curriculum : (a) Step 0−200k: https://github.com/facebookresearch/fairseq/tree/nllb/examples/ nllb/modeling/scripts/flores200/cl1_lang_pairs.txt (b) Step 200k−300k: https://github.com/facebookresearch/fairseq/tree/nllb/examples/ nllb/modeling/scripts/flores200/la...

    2021

Pith tools

Reviewed May 12, 2026 · model on record in the stance chip above.